Calibration Status Classifier¶

The Calibration Status Classifier is a camera-LiDAR binary classification model for calibration health monitoring. It predicts whether the current sensor calibration is valid or whether recalibration is required.

Summary¶

Property	Value
Task	Calibration status classification
Modality	Camera and LiDAR
Input	Fused five-channel BGRDI image
Output	Binary calibration status
Architecture	ResNet18 backbone with global average pooling and head
Datasets	NuScenes, T4Dataset

Available Configurations¶

Experiment configs live under autoware_ml/configs/tasks/calibration_status/calibration_status_classifier/ and are passed to the CLI as the path relative to tasks/.

Config Name	Dataset	Purpose
`calibration_status/calibration_status_classifier/nuscenes/resnet18_b8_30e_nuscenes`	NuScenes	Standard training configuration
`calibration_status/calibration_status_classifier/t4dataset/base`	T4Dataset	Standard training configuration

The T4Dataset row is a base. A T4Dataset experiment binds it together with its database in an external config root, see Configuration.

Input Representation¶

The model consumes a fused image with five channels.

Channel	Content
`0-2`	BGR image
`3`	LiDAR depth projection
`4`	LiDAR intensity projection

Training¶

autoware-ml train --config-name calibration_status/calibration_status_classifier/nuscenes/resnet18_b8_30e_nuscenes

For a pipeline validation run:

autoware-ml train \
    --config-name calibration_status/calibration_status_classifier/nuscenes/resnet18_b8_30e_nuscenes \
    +trainer.fast_dev_run=true

Evaluation¶

autoware-ml test \
    --config-name calibration_status/calibration_status_classifier/nuscenes/resnet18_b8_30e_nuscenes \
    --weights mlruns/calibration_status/NuScenesDataset-base-v1.0.0/nuscenes/resnet18_b8_30e_nuscenes/<run_id>/artifacts/checkpoints/best.ckpt

Deployment¶

autoware-ml deploy \
    --config-name calibration_status/calibration_status_classifier/nuscenes/resnet18_b8_30e_nuscenes \
    --weights mlruns/calibration_status/NuScenesDataset-base-v1.0.0/nuscenes/resnet18_b8_30e_nuscenes/<run_id>/artifacts/checkpoints/best.ckpt

The exported model keeps the original calibration-status behavior and emits a single probability tensor. Class labels can be derived downstream with argmax when needed.

Data Pipeline¶

The training pipeline includes image undistortion, synthetic calibration perturbation, LiDAR-camera fusion, and channel-first tensor conversion. Calibration perturbation is used to generate positive training samples representing miscalibrated sensor pairs.

Implementation¶

Path	Description
`autoware_ml/models/calibration_status/`	Model implementation
`autoware_ml/datamodule/t4dataset/dataset.py`	T4Dataset records as a dataset
`autoware_ml/datamodule/nuscenes/dataset.py`	nuScenes dataset
`autoware_ml/transforms/camera_lidar/camera_lidar.py`	Calibration perturbation and LiDAR-camera fusion
`autoware_ml/configs/datamodule/`	Datamodule, dataset and transform configurations
`autoware_ml/configs/tasks/calibration_status/calibration_status_classifier/`	Experiment configurations